GitHub Copilot is $10/month. Cursor is $20/month. Meanwhile three open-source projects claim to do the same job for $0. I don't trust claims, so I installed all three and used each as my only AI coding assistant for a full week.
Same machine, same projects (a FastAPI backend + a React frontend), same local model backend (Ollama + qwen2.5-coder:7b). Here's what actually happened.
The Contenders
| Continue | Tabby | MonkeyCode | |
|---|---|---|---|
| Price | Free | Free (self-host) | Free |
| GitHub stars | ~26K | ~30K | Growing |
| Setup time | 5 min | 40 min 😬 | 5 min |
| IDE support | VS Code, JetBrains | VS Code, JetBrains, Vim | VS Code, JetBrains |
| Backend flexibility | Any (Ollama, OpenAI, etc.) | Built-in server | Any OpenAI-compatible |
| Completion quality (my tests) | Good | Medium | Good |
| Chat + inline edit | ✅ | ⚠️ basic | ✅ |
Week-by-Week Notes
Week 1 — Continue. The config file (~/.continue/config.json) is both its superpower and its curse. You can wire up anything, but you'll spend an hour in JSON first. Tab completion worked well; the @codebase context feature genuinely helped on a 40-file project.
Week 2 — Tabby. I wanted to love it — fully self-hosted, your own completion server, great privacy story. But the setup hurt: Docker compose, model registry quirks, and the completion quality with my 7B model was noticeably weaker (more partial-line garbage suggestions). If you have a GPU server and an ops mindset, maybe. For a laptop dev, too much friction.
Week 3 — MonkeyCode. Installed, pointed it at my local Ollama endpoint, done. The inline-edit flow (select code → describe change) was the smoothest of the three, and it handled the "explain this legacy function" case better than I expected from a free tool.
By the Numbers (My Usage Logged for 3 Weeks)
| Metric | Continue | Tabby | MonkeyCode |
|---|---|---|---|
| Suggestions accepted | 61% | 44% | 68% |
| Avg. completion latency | 2.1s | 2.9s | 1.9s |
| Config/setup pain (1-10) | 4 | 8 | 2 |
| Times I reached for Copilot | 3 | 7 | 1 |
The Controversial Take
Self-hosting is not automatically better. Tabby's pitch — "run your own Copilot server" — sounds ideologically pure, but if the setup costs you a weekend and the suggestions are 25% worse, you haven't beaten the $10 subscription. You've just paid it in time instead of money. The best free tool is the one you'll actually keep installed.
My ranking for a solo dev on a laptop: MonkeyCode > Continue > Tabby. If you're on a team with a GPU box and a DevOps person, re-run the experiment — Tabby probably climbs.
Try them yourself — MonkeyCode is here: https://ly.cyberserval.tech/iIETXiF (free, open-source, works with Ollama out of the box).
Have you run a similar bake-off? What acceptance rates are you seeing with local models — and does anyone actually still pay for Copilot after trying the free tier of alternatives?
Top comments (0)